Talk:Ascendant/RRGC

From Robowiki
Revision as of 16:26, 24 April 2009 by Voidious (talk | contribs) (migrating from old wiki)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search

1889 points! This must be the strongest gun ever. -- PEZ

Yeah. I already feared that there is something wrong with my gun testbed, since the new gun performed even better than Bee there. Now i'm tempted to plug this gun into Ascendant as it is.

BTW: I tested alot with Bee to see why it performed better than my earlier test versions. It turned out to be your wall segmentation which is much better than to simply segment on distance to wall by using the targets current heading. So i'm now using your type of wall segmentation (with different granularity), so thanks&credits to you PEZ. --Mue

But it performs much better than Bee in the RRGC too, so what's wrong with your test bed? Cool that you found my wall segmentation being useful! I've said it a few times on this wiki that I think it's a very good segmentation. What granularity are you using then? Did you lower it slightly? Mine is tweaked using the TargetingChallenge, which I think is a bad testbed for this. I'll try using a lower granularity now I think. -- PEZ

No, there is nothing wrong with my testbed apparently. I just did not trust it that much, since i did not really expect the new gun to do better than Bee (which has been tuned in a lot of releases in the rumble). To granularity of wall segments: I use 47 guess factors and 5 wall segments with wall segment width being 7 guess factor bins. --Mue

You've got 10 points up to Shadow. Your new gun performs some 17 points better than your old. You just might make it to the throne again with this gun. You splice the wall segmentation up in GF bins? That's pretty clever. -- PEZ

  • That was the way i read your code in Bee, so its not my cleverness this time :-) --Mue

BTW, blessing your gun with my wall segmentation boosts it's performance well past my gun. What do you think I am lacking in my gun now? -- PEZ

Hm, i dont really know. My first guess would be that rolling stats work better. Apart from that maybe your bullet power segmentation costs you some points. --Mue

Rolling stats. I'll try that. Hopefully you're right about the bullet power segmentation slowing things down. I've removed it in the latest release. -- PEZ

Which bots do you have in your gun testbed? I've experimented some with rolling the stats, but it doesn't seem to improve my gun. Removing bullet power segmentation lost me RR@H points. Any other ideas? -- PEZ

My testbed for targeting experiments consists of:

Dont know what else it could be that is holding Bee back. I use acceleration segments instead of last velocity segments (need less segments this way), but i dont think that this has an impact on performance. --Mue

But it just might have an impact on performance. At least it might mean that I do not benefit much from using two velocity segments. Which I thought I did. Time to experiment and question everything in my gun I guess. -- PEZ

I'm impressed! You managed to press another 10 points out of your already excellent gun. Now, please release an Ascendant using this gun. I think your movement already is almost on par with Shadow's. Now with a gun that is clearly betterm who's going to be King? =) -- PEZ

Yep, I can't wait to see Ascendant equipped with this gun, why the wait? And what about some TC scores? -- ABC

I just didnt want to put an unfinished gun into version 1.0 :-). Last on my todo list is some bullet power tweaking. But since i dont really expect this to bring me any points, i'll take the gun as it is now and put it in a release of Ascendant tomorrow - i'm curious too :-). I'll also take a look at the TC then, though i consider the RRGC score to be a better measure for targeting performance (at least with respect to the rumble). --Mue

Yes, the RRGC is a far better measure. But even so, I'm curious what this beast does in the TC. -- PEZ

I like the TC because it measures long term performance. -- ABC

I like gut instinct because I have no patience. :-) --David Alves



Yay, gut instinct! My favorite way of testing, too. Maybe that's the problem with us Americans... -- Alcatraz

Tis' nothing wrong with gut feeling other than that it's almost always wrong. =) How about some TC500 results Mue? I think that's what ABC means with "long term performance". -- PEZ

Check the TC500 again, Ascendant is in 1st place. :) -- ABC

Yes, I missed it on the Changes listing. Only the slightest margin to CC though. The margin of error is larger than that. From my biased position it's a draw. =) -- PEZ

It's a draw from my (unbiased) position too, but the TCFast and RRGC tell a different story... -- ABC

I'm working on that. I'm working on that. -- PEZ


Yeah, you are right, its a draw. I suspect that the fast learning buffer somehow lowers performance in the long run. I'll experiment with that a bit. --Mue

I'd be happy to settle the draw by running 10 seasons of TC500 for each of them if you send me non-moving versions of your the bots. --David Alves

I'm not sure I have the exact version of Bee left around... But maybe the new one can compete. I'll run one TC500 and check. -- PEZ

I think I have CC 1.9.9.87 already... Mue, if you send me a TC version of Ascendant I'll run each of them through 10 rounds of TC500 and settle this once and for all. :-) --David Alves

Edit conflict... OK. It seems the new Bee can perform like that old one. 94.00 in my TC500 test. And during that test SandboxDT was too slow to start 8 time and TheArtOfWar one time which made me lose some in score I guess. Still the margin to Ascendant could well be within the margin of error so I'd welcome your test, David. Download the TC-version of CC here: http://pezius.com/robocode/robots/pez.rumble.CassiusClay_1.9.9.87TC.jar -- PEZ

10 rounds of TC500? That would take like 8 hours on my machine and another 8h for CC i think. Thats a lot of computation time ... Anyway, i uploaded a TC version of Ascendant to: http://informatik.hu-berlin.de/~ueckerdt/mue.Ascendant_1.0TC.jar. Please tell if you cant reach this domain (i know PEZ couldnt, but i have no other web space at hand). --Mue

They're running now. I'll upload the results when I wake up tomorrow. May the best bot win! --David Alves

Results posted. --David Alves

Way cool! I was pretty confident I would win when I saw my solid #1 score in TC35. So now I have my TC crowns back. Why is my gun not competing with Mue's at all in the RR@H? A Riddle with capital R... -- PEZ

Ding dong! 1.0.2 is out. What's up for test now? -- PEZ

Ah someone is watching :-). I just had no time to add a description, but added one now. --Mue

I'm totally tuned in on the Ascendant Gun development channel! My tests with segmenting on reverse wall proximity hasn't hinted at a success. But you seem to at least retain your excellent score... Maybe I should try a release using that segmentation. What's your wall segmentaion tweak about? -- PEZ

My tests didnt really show an improvement either, i just released it anyway :-). The other 'wall segmentation tweak' was a small code change that allows me to use wall segments of different size. So the length of the gf intervals covered by the different wall segments may vary from segment to segment. I thought up some numbers there, lost patience while testing and released it... :-) --Mue

Yes, maybe a non-linear distribution of wall-segment indexes would work. Worth a try! -- PEZ

It looks like 1.0.4 is very strong against surfers. 45+% against Shadow... FYI I refactored my wall segmentation so that I now can choose the segment index widths at will. I think I'll make the same refactoring for all my segmentations. (Even if for the wall segmentatuion I haven't found a setup that works better than the linear distribution I already had.) -- PEZ

Maybe its something special with Shadow, since i dont have an idea why the current wall segmentation should be better against surfers (i only changed segment widths). And i've also come to the conclusion that different segment widths dont really help with wall segmentation. There was just no reason to assume this when i started experimenting.

To this refactoring: I use array constants that contain the segment borders (probably the same way you do it). This makes it very easy to experiment with segmentation since i only need to modify the array in order to add, remove or resize the segments. In addition to that its very easy to see what segmenation is currently used (i still remember the time-since-velocity-changed segmenation in Bee where i gave up to figure out what segmentation borders you actually use :-). --Mue

Yes, I use literal array constants too and benefit from the same effects of it as you do. =) That timer segmentation formula in Bee was developed using Excel b t w. But I'll change to use array constants instead since it'll allow hand tweaking with much greater ease. -- PEZ

Man, you're an even tougher game than Jamougha ever was! 1904. Care to toss me a hint on what your latest segmentation tweak was about? -- PEZ

Thanks :-), looks like the first place of RRGC has come in reach. The tweak was still about reverse wall proximity. I'm now using 3 segments, each covering 0.4*escapeAngle. --Mue

I'll change my wall segmentation semantics to use fractions of the escapeAngle too. That's so much better and how I should have done it to begin with. Thanks! -- PEZ

Retrieved from "http://robowiki.net/w/index.php?title=Talk:Ascendant/RRGC&oldid=5448"